Animals don't give off much waste heat. They evolved not to. And it's very hard to extract energy from small temperature gradients. The thermoelectric chargers that convert body heat to energy are only ~1% efficient. Compare to sunlight, which gives ~20x the watts per square inch, and can be converted at 20% efficiency. So a body heat charger needs to be roughly 400 times as big as a solar panel does.
This doesn't seem right. Anyone who has been in a cattle barn knows the kind of heat that cows give off. The energy demand for the sensors should be very small. 20% efficiency for the solar cells might be true on paper, but not in practice - diet, orientation, even needing 4x the solar cells because you cant guarantee the cells will always face up (collar rotates).
He originally vibe coded this as AI Exposure, but deleted that one because people were misinterpreting it. Someone mirrored that one here https://joshkale.github.io/jobs/ . EDIT: D'oh, I now see Karpathy didn't delete it, he just made it into a "Digital AI Exposure" button.
Surely I am misreading what you mean here. But your comment very much reads like you think "Iran strikes Israel on March 10" at 25% odds is what in fact caused Iran to strike Israel on March 10. I'll search the space of other interpretations in the mean time, but if you could help me out here and clarify…
> But your comment very much reads like you think "Iran strikes Israel on March 10" at 25% odds is what in fact caused Iran to strike Israel on March 10
I don't know how you extrapolated that from the parent's comment. It literally said nothing about the cause and effect of this particular event.
Knowing the odds in a prediction market IS a big part of the problem brought up in the linked article though (and the bets themselves). Knowing how much can be made from being right creates an upper-bound on what a financially-rational malicious actor will spend in trying to change the outcome.
Their comment proposed something that would "be the answer here".
What does "here" mean? It's logical to expect "here" to refer to a scenario that includes cases like the one in the article. If it's some scenario that excludes cases like the one in the article, then it's not actually relevant to the discussion.
(Tangents are OK. It's just confusing if they're introduced with phrasing that makes them sound like they're not tangents.)
"here" in that comment is not referring to any specific scenario. It is referring to the problem discussed in the sentence immediately following it, that public prediction markets can shape the outcome of the events they are predicting.
I wasn't extrapolating - it was the literal meaning of the words. The context was that someone commented "shouldn't be called prediction markets, they should be called outcome-shaping markets" in direct reply to "Polymarket gamblers threaten to kill me over…[the prediction market "Iran strikes Israel on March 10"]. I interpret that as polymarket gamblers outcome-shaped Iran striking Israel. It was at 25% odds when they struck. I don't think the commenter actually meant that literally, which is why I asked them to clarify. I'm just doing my best here.
I'm not even talking about the specific situation. The problem with prediction markets that has been talked about for months is the predictions themselves are being used as an indicator that "thing will happen" and eventually there is so much liquidity on certain markets that the market determines the outcome not the other way around
> I'm not parent commenter, but any example of an event that actually matters?
Elections, wars, the economy, healthcare systems, laws, court cases?
I think the point is more that if it's already happening at this level of adoption, and these are only the ones we know about, surely there are many more where this is happening and we don't know, and as adoption increases, it will get worse and worse since the liquidity on the line will be much higher.
I'm not parent commenter, but any example of an event that actually matters?
Elections, wars, the economy, healthcare systems, laws, court cases?
The example you provided does look like some insider trading type of stuff, just like how 1v1 professional sports players will sometimes intentionally lose after receiving a bribe to do so, but both your example and sports don't really seem to have any kind of meaningful impact on anything or anyone who isn't gambling, no?
I don't know. I imagine if someone was going to adjust the timing or targets of strikes in a war in response to a prediction market bet, or the decisions of a high-profile court case etc, they wouldn't say so in public the way the CEO of Coinbase did on that earnings call. It's pretty rare that someone would actually claim they're taking an action specifically with the purpose of altering the outcome of the prediction market bet, rather than giving some other reason. So even if it were happening, the only evidence I might expect to find would be suspicious prediction-market trades around low-probability events.
In situations like this, where you wouldn't expect to see evidence of something even if it were happening, you're basically left to make a judgement based on prior probability. So here that would come down to: is the financial incentive provided by the prediction market high enough relative to the decisionmaker's perceived risk and penalty of being caught? IMO, the answer is currently 'no' for most high-profile cases, but in a future where more money is riding on the bets, or where decisionmakers are insulated from consequences, that could swing to 'yes'.
The decision-makers don't need change their decision whatsoever to corruptly profit from it, though, they can just place bets on the timeframe or outcome of whatever decision they originally intended to make. Why risk operational disruptions and a greater chance of getting caught when you can still profit exorbitantly from insider trading on your knowledge without incurring those unnecessary risks?
How can we possibly find an example, if the names of the betters aren't public? There are public bets on activities from the IS government that can easily be used by people that control the outcome.
Maybe not directly so clearly, but there is some influence factor for sure. For example we see this with sports betting, at the far end of the spectrum it is literally players or coaches fixing games to satisfy bets. But somewhere in between that overt fraud, there is influence making going on. Submarine stories on ESPN or sports blogs highlighting a players bad shoulder, hurting their perceived value going into a free agency period.
This applies to governance as well. Note how there was a bet placed on polymarket on maduros capture mere hours before the raid were conducted, and this has lead to legislation moving forward in effort to combat suspicious prediction market activity based on whitehouse insider knowledge.
I was confused as well: After further consideration, I realized that Polymarket only resolves on actual event outcomes insofar as mainstream journalism accurately and credibly reports on those events and their resolution criteria rely on those reports. When it's easier or more reliable to influence reporting around outcomes than the outcomes themselves, bettors will seek to influence the behavior of reporters rather than the behavior of eg. national militaries.
"Iran strikes Israel on March 10" is a difficult outcome to force one way or the other. But "The Times of Israel’s military correspondent or other credible sources reports that Iran strikes Israel on March 10" merely requires intimidating or bribing one journalist. The existence of the bet didn't cause the missile strike or the failed interception. But it did cause a significant, heavily motivated dis-information campaign from people who stood to lose a lot of money.
In a similar way, people fear that prediction markets estimating times of death are equivalent to assassination markets. But murder is an aggressively prosecuted serious crime. It seems that it would be far easier and lower risk to bribe, coerce, pretend to be, or literally be a reporter who got a false obituary published - wait until the "victim" is going to be offline for a few days and can't be contacted to prove they're not dead, trick some tropical country's coast guard into confirming that the victim's yacht exploded offshore, point Polymarket at the obituaries, grab all the crypto, and disappear. If you fail to disappear successfully, your worst crime is publishing a fictional news article/bribing/threatening a journalist. You don't have to risk being in the legal jurisdiction of the victim and getting your hands bloody, you don't even have to be in the same hemisphere: you only need convince the prediction market resolution criteria that something happened.
Scott Alexander wrote about a related issue last month in the colorfully named section [1] "Annals of the Rulescucks", where he described a half dozen scenarios where the outcome may or may not have diverged from the actual event. A bet isn't resolved by eyewitnesses, it's resolved over the Internet through another financial instrument.
> you only need convince the prediction market resolution criteria that something happened.
You don't even need the market to actually resolve, as soon as your rumor becomes credible the market pricing will adjust and you can just sell your shares before resolution.
I guess the next step in this evolution is to set up controlled news sources. You get people who have an official press card to report on things as you need as part of the reporting manipulation business.
"Hey there's this newspaper that says this obscure thing happened, please resolve the bet in my favour"
It doesn't sound difficult to solve. The sensors can classify firetruck, ambulance, red&blues, uniformed police, badge, siren, etc. At a certain criteria, they can unlock the driver door for normal human driving, perhaps for a very limited speed and distance. The officer can move it to the side, and if they crash it's not waymo crashing it. This override should send the an alert to the remote command center, so a human can watch the video and also decide how much further they can drive it. Since passenger safety is a concern, if there is a passenger inside who chooses to remain inside, the car should remain locked and not allow any driver in. The human can decide to follow police orders to exit the car, or remain inside, but at that point a human becomes responsible for obstructing. The whole freezing waymo trend seems driven by legal liability - not engineering. They know if they always freeze, their million miles with no accidents stats are safe.
> Cops can't move vehicles that they don't own because of liability. The only way for them to move a vehicle without liability is to use a tow truck.
While the precise boundaries of liability depend on the laws of the particular jurisdiction (they aren't consistent across the whole US) police generally can take reasonable action to move vehicles obstructing the road in an emergency without liability for any damages incurred, whether or not they use a tow truck to do it.
I mean liability can be defined, we are writing new laws for these things.
Cops that need to move autonomous are not liable for any damages.
I do think that theres needs to be better handling of emergency vehicles and autonomous vehicles. Someone needs to spend some serious time thinking through how to handle this better because this situation was not okay.
They are great, but they are expensive. I can run those against the cheaper ollama cloud models for things that are basically requirements gathering and review of a plan. The Product Designer Agent and the Product Manager basically argue for a few rounds and give an artifact that the coding agents pick up.
It could all easily be anthropic models and would work well, but running this swarm eats up all my anthropic tokens and these other models are good enough for the roles I've given them.
It's too slow. It takes at least five seconds to load the next picture after you answer. You should probably just preload all the pictures client-side. I wasn't able to get through it.
That's not what preload is for. You don't wait for the extra images on the first page. You start loading them after the page is complete, so that the next page loads faster.
I'm not sure there was a technical way to preload back then.
I wasn't working in web dev in the early 00s, but I had done some informally. I just don't think preloading was a thing.
It's been a long time, but I'm pretty sure AJAX does not predate this website, SPAs were impossible, etc. I suppose you might do something with IFRAMEs, but you likely wouldn't be able to fetch those /after/ the main page. Preloading would be simul-loading and lag the hell out of your main window.
The way to instruct a browser to load back then was to wait for the user to actually click to fetch, and that would navigate away from the current page.
Yes. The tolerance for slow-loading pages has shrunk dramatically since the early days of the web. Admittedly, 20 seconds was pushing it and people tried to reduce image sizes and such things to make it faster. But I recall in the late 90s, when I was selling CMS software to IBM, they had a rule that all pages needed to load within 3 seconds, and that was considered fast in those days.
reply